Paper Style Review

结构化审校中文论文：从参考论文学习风格，统一结构图谱，执行格式/术语/逻辑/风格融合审查，并将结果回写为 Word 批注。

Security Scan

The skill's code and README show it needs an LLM API key and will read/write local Word documents, but the published metadata omits required credentials and there are runtime path-inference behaviors that are not fully explained — these mismatches warrant caution before installing.

Purpose & Capability

The name/description and included scripts (document parsing, style profiling, format checking, and Word comment injection) are coherent with a paper-style-review tool. However, the registry metadata declares no required environment variables or credentials, while README/SKILL.md and the code refer to LLM configuration (e.g., PAPER_STYLE_REVIEW_LLM_API_KEY or OPENAI_API_KEY) and expect an OpenAI-compatible endpoint. Omitting these required credentials in the manifest is an incoherence that could mislead users about what the skill needs.

Instruction Scope

SKILL.md clearly limits scope (no embedded reference corpora, target-stage should only read style-profile.json, etc.). The scripts implement the advertised pipeline. Still, some runtime behaviors merit attention: helper code (annotation_anchor_infra.infer_target_path) will try to infer target docx paths by reading run-summary.json or by traversing parents to find a file named '论文修改版.docx' — this can cause the skill to read files outside the explicit output/input paths if placed in certain workspaces. The README also indicates the skill will read user-provided refs and target .docx files and call an LLM; those runtime reads/calls are expected, but the path-inference heuristics are not documented in SKILL.md and could access unexpected local files.

Install Mechanism

There is no install spec in the registry (instruction-only), but the skill bundle includes many runnable Python scripts. That means installation is a matter of running the included Python code and installing its Python dependencies (openai, python-docx, lxml, PyMuPDF, etc.) — no external binary downloads were observed in the provided files. This is a moderate risk: code will execute locally, but there are no opaque remote install steps in the package itself.

Credentials

Although the registry lists no required environment variables, the README and code reference LLM-related environment variables (PAPER_STYLE_REVIEW_LLM_API_BASE, PAPER_STYLE_REVIEW_LLM_API_KEY, PAPER_STYLE_REVIEW_LLM_MODEL and compatible OPENAI_* variables). Requesting an LLM API key is proportionate for functionality, but the manifest omission is inconsistent and could hide expected network access. Additionally, some config options mention external term validation (CNKI) and local terminology paths; those features are optional but increase the places where secrets/config might be needed. The code also inspects workspace paths (see infer_target_path), which has privacy implications for local credentials or documents present in the workspace.

Persistence & Privilege

The skill does not request 'always:true' and does not declare persistent elevated privileges. It is runnable on demand (user-invocable). There is no evidence in the provided files that it modifies other skills' configs or system-wide agent settings. Autonomous invocation is enabled (default) — normal for skills — but combined with the missing-declared-credentials issue this increases the importance of carefully scoping API keys used by the skill.

What to consider before installing

This package largely implements what it says (parsing .docx, deriving a style profile from user-supplied refs, running format/logic/style checks, and writing Word comments), but there are important mismatches and operational details to check before use: - The manifest lists no required env vars, but the README and code expect an LLM API key (examples: PAPER_STYLE_REVIEW_LLM_API_KEY or OPENAI_API_KEY and an API base). Expect the scripts to call an external LLM endpoint; do not run with your primary org key unless you are comfortable with that provider and network traffic. Consider creating a dedicated, rate-limited API key. - The code will read/write local .docx files you point it at. However, helper code contains heuristics (e.g., looking for run-summary.json or a file named '论文修改版.docx' in parent directories) that could read files outside the explicit inputs/outputs. Run the pipeline from a controlled working directory and inspect infer_target_path and related functions if you want to be certain which files will be accessed. - There is no remote installer, but you must install Python dependencies (openai, python-docx, lxml, PyMuPDF, etc.). Review requirements and run in a virtualenv or container. - Before installing or running: inspect llm_client.py to confirm which endpoints are called, and confirm no unexpected external endpoints beyond your chosen LLM provider. If you need stronger guarantees, sandbox execution (isolated VM/container) and use of a dedicated API key are recommended. - If the publisher updates the registry metadata to explicitly declare required env vars and documents the path-inference behavior, that would reduce ambiguity. If you can obtain such clarifications (or a signed manifest that includes the expected env vars), the assessment could be upgraded. In short: functionality matches purpose, but metadata/documentation mismatches and workspace path heuristics are reasons to treat this skill with caution and to run it in an isolated environment with dedicated API credentials.

用这个 skill 处理中文论文的“风格学习 + 结构化审校 + Word 批注回挂”任务。

保持的公开版边界

• 只保留可公开发布的脚本、schema、派生规则和运行说明。
• 不保留真实论文、命名样例、历史输出、原始 .doc 附件或本地工作区绝对路径。
• 风格学习依赖用户自行提供的 refs，仓库不内置参考论文。

当前主链

scripts/run_review.py
  -> orchestrators/review_orchestrator.run_review()
  -> style-profile handoff or build
  -> format checker
  -> target style units
  -> fused review runtime
  -> annotation assembly
  -> inject Word comments

能力范围

1. refs 风格学习

   • 为每篇 ref 建立统一结构图谱
   • 按 paragraphType 聚类
   • 生成 style-profile.json 与 ref-style-basis.md

2. target 审查

   • 建立 target 结构图谱
   • 运行格式检查
   • 运行 fused LLM 审查：术语 / 逻辑 / 风格偏离

3. Word 批注回挂

   • 汇总为 annotations.json
   • 生成 <target>-annotated.docx

关键边界

• target 风格判断只消费 style-profile.json，不在 target 阶段回读 refs 原文。
• 风格学习、target 审查、批注定位共用统一结构图谱。
• 所有正式输出都走统一锚点与注入链路。
• chapterScope 只收窄 target-side LLM 审查；格式检查仍是文档级。

运行方式

完整主链：

python3 scripts/run_review.py --config ./references/review-config.example.json

独立阶段：

python3 scripts/run_review.py --config ./references/review-config.example.json --stage style_profile
python3 scripts/run_review.py --config ./references/review-config.example.json --stage style_annotation
python3 scripts/run_review.py --config ./references/review-config.example.json --stage annotation

限制章节范围：

python3 scripts/run_review.py --config ./references/review-config.example.json --chapters abstract,1

最小配置

{
  "target": {
    "path": "./inputs/target.docx"
  },
  "refs": [
    {
      "id": "ref-style-01",
      "path": "./inputs/ref-style-01.docx",
      "enabled": true,
      "styleProfileEnabled": true
    }
  ],
  "outputDir": "./outputs"
}

重要文件

• scripts/job_defaults.py / scripts/runtime_config.py：配置归一化
• scripts/paper_structure_anchor.py：统一结构图谱
• scripts/profilers/ref_style_profiler.py：refs 风格学习
• scripts/fused_review_runtime.py：target-side fused 审查
• scripts/annotation_anchor_infra.py / scripts/annotation_assembler.py / scripts/inject_word_comments.py：批注链路
• references/format-rules.json / references/format-rule-matrix.md：公开可分发的派生格式规则
• references/style-profile-extractor.md：风格画像提取约束

使用时要做的事

1. 准备自己的 target 与 refs。
2. 按配置文件指定 outputDir、检查项和可选 chapterScope。
3. 若已存在稳定 style-profile.json，可直接通过 styleProfilePath 复用。
4. 调试结构或定位问题时，优先检查 target-structure-map.json、target-style-units.json、annotations.json。

不要做的事

• 不要假设仓库内自带任何真实论文样本。
• 不要把 target 阶段改成直接回读 refs 原文。
• 不要在公开版里加入历史输出、用户文档、原始附件或本地路径。